CONTROL AND DYNAMIC SYSTEMS
Advances in Theory and Applications Volume 71
CONTRIB UTORS TO THIS VOLUME TONGWEN CHEN ...
12 downloads
479 Views
10MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
CONTROL AND DYNAMIC SYSTEMS
Advances in Theory and Applications Volume 71
CONTRIB UTORS TO THIS VOLUME TONGWEN CHEN PA TRIZIO COLANERI OSWALDO L. V. COSTA PA OLO D ' ALESSANDR O ELENA DE SANTIS BRUCE A. FRANCIS JEAN- CLAUDE HENNE T SHERWOOD TIFFANY HOADLEY O. THOMAS HOLLAND VIVEK M UKHOPADHYA Y WENDY L. POSTON A N T H O N Y S. POTO TZKY CAREY E. PRIEBE RICCARDO SCA TTOLINI NICOLA SCHIA VONI H A N N U T. TOIVONEN CAROL D. WIESEMAN
CONTROL AND DYNAMIC SYSTEMS ADVANCES IN THEORY AND APPLICATIONS
Edited by
C. T. LEONDES School of Engineering and Applied Science University of California, Los Angeles Los Angeles, California
V O L U M E 71:
DISCRETE-TIME CONTROL SYSTEM ANALYSIS AND DESIGN
ACADEMIC PRESS San Diego New York Boston London Sydney Tokyo Toronto
This book is printed on acid-free paper. Q
Copyright 9 1995 by ACADEMIC PRESS, INC. All Rights Reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. A c a d e m i c Press, Inc. A Division of Harcourt Brace & Company 525 B Street, Suite 1900, San Diego, California 92101-4495
United Kingdom Edition published by Academic Press Limited 24-28 Oval Road, London NW1 7DX
International Standard Serial Number: 0090-5267 International Standard Book Number: 0-12-01277 I-7
PRINTED IN THE UNITED STATES OF AMERICA 95 96 97 98 99 00 QW 9 8 7 6
5
4
3
2
1
CONTENTS CONTRIBUTORS .................................................................................. PREFACE ................................................................................................
vii ix
H2-Optimal Control of Discrete-Time and Sampled-Data Systems
Tongwen Chen and Bruce A. Francis Techniques for Reachability in Input Constrained Discrete Time Linear Systems ........................................................................................
35
Paolo d'Allesandro and Elena De Santis Stabilization, Regulation, and Optimization of Multirate Sampled-Data Systems ............................................................................
95
Patrizio Colaneri, Riccardo Scattolini, and Nicola Schiavoni Maximizing the Fisher Information Matrix in Discrete-Time Systems
.... 131
Wendy L. Poston, Carey E. Priebe, and O. Thomas Holland Discrete Time Constrained Linear Systems ............................................ 157
Jean-Claude Hennet Digital Control with H~ Optimality Criteria
........................................... 215
Hannu T. Toivonen Techniques in On-Line Performance Evaluation of Multiloop Digital Control Systems and Their Application .................................................. 263
Carol D. Wieseman, Vivek Mukhopadhyay, Sherwood Tiffany Hoadley, and Anthony S. Pototzky
CONTENTS
vi
Impulse Control of Piecewise Deterministic Systems ............................
29 1
OmValdo L. V . Costu
INDEX .....................................................................................................
345
CONTRIBUTORS Numbers in parentheses indicate the pages on which the authors" contributions begin.
Tongwen Chen (1), Department of Electrical and Computer Engineering, University of Calgary, Calgary, Alberta, Canada T2N 1N4 Patrizio Colaneri (95), Dipartimento di Elettronica e Informazione, Politecnico di Milano, 32-20133 Milano, Italy Oswaldo L. V. Costa (291), Department of Ele~'tronics Engineering, Escola Politfca da Universidade de Sfio Paulo, 05508 900 S(to Paulo SP, Brazil Paolo d'Alessandro (35), Department of Mathematics, 3rd University of Roma, 00146 Roma, Italy Elena De Santis (35), Department of Electrical Engineering, University of L'Aquila, 67040 Poggio di Roio, L'Aquila, Italy Bruce A. Francis (1), Department of Electrical Engineering, University of Toronto, Toronto, Ontario, Canada M58 1A4 Jean-Claude Hennet (157), Laboratoire d'Automatique et d'Analyse des Systomes du CNRS, F-31077 Toulouse, France Sherwood Tiffany Hoadley (263), Langley Research Center, National Aeronautics and Space Administration, Hampton, Virginia 23681 O. Thomas Holland (131), Naval St, face Walfare Center, Dahlgren Division, Dahlgren, Virginia 22448 Vivek Mukhopadhyay (263), Langley Research Center, National Aeronautics and Space Administration, Hampton, Virginia 23681
vii
viii
CONTRIBUTORS
Wendy L. Poston ( 131 ), Naval Walfare Center, Dahlgren Di~'ision, Dahlgren, Virginia 22448 Anthony S. Pototzky (263), Lockheed Engineering and Science Co., Hampton, Virginia 23666 Carey E. Priebe (131), Naval Sulfate Warfare Center, Dahigren Division, Dahlgren, Virginia 22448 Riccardo Scattolini (95), Dipartimento di Elettronica e lnformazione, Politecnico di Milano, 32-20133 Milano, Italy Nicola Schiavoni (95), Dipartimento di Elettronica e Informazione, Politecnico di Milano, 32-20133 Milano, Italy Hannu T. Toivonen (215), Process Control Laboratory, Depap'tment of Chemical Engineering, Abo Akademi Uni~'ersitv, 20500 Turku/Abo, Finland Carol D. Wieseman (263), Langley Research Center, National Aeronautics and Space Administration, Hampton, Virginia 23681
PREFACE Effective control concepts and applications date back over millennia. One very familiar example of this is the windmill. It was designed to derive maximum benefit from windflow, a simple but highly effective optimization technique. Harold Hazen's 1932 paper in the Journal of the Franklin Institute was one of the earlier reference points wherein an analytical framework for modem control theory was established. There were many other notable landmarks along the way, including the MIT Radiation Laboratory Series volume on servomechanisms, the Brown and Campbell book, Principles of Servomechanisms, and Bode's book entitled Nem'ork Analysis and Syntheses Techniques, all published shortly after mid-1945. However, it remained for Kalman's papers of the late 1950s (wherein a foundation for modem state space techniques was established) and the tremendous evolution of digital computer technology (which was underpinned by the continuous giant advances in integrated electronics) for truly powerful control systems techniques for increasingly complex systems to be developed. Today we can look forward to a future that is rich in possibilities in many areas of major significance, including manufacturing systems, electric power systems, robotics, and aerospace systems, as well as many other systems with significant economic, safety, cost, and reliability implications. Thus, this volume is devoted to the most timely theme of "Discrete-Time System Analysis and Design Techniques." The first contribution to this volume is "H2-Optimal Control of DiscreteTime and Sampled-Data Systems," by Tongwen Chen and Bruce A. Francis. This contribution presents a state space solution to the discrete He control problem and also presents direct formulas for an H2-optimal sampled-data control problem with state feedback and disturbance feedforward. The derivations presented in this contribution are new and quite self-contained, with the derived formulas being applicable to the sampled-data problem via the powerful lifting technique, which is described in the latter part of this chapter. As such, this is a most important contribution with which to begin this volume. The next contribution is "Techniques for Reachability in Input Constrained Discrete Time Linear Systems," by Paolo d'Alessandro and Elena
x
PREFACE
De Santis. Constraints on the input of a discrete-time system result in constraints (reachability) on the system state. Therefore, this issue is of essential importance in the analysis and design of discrete-time systems. This contribution is an in-depth treatment of the many aspects involved in this essential issue. The next contribution is ~'Stabilization, Regulation, and Optimization of Multirate Sampled-Data Systems," by Patrizio Colaneri, Riccardo Scattolini, and Nicola Schiavoni. There are two primary reasons for the importance of multirate digital control in practice. One of these is the fact that, in practice in many diverse applications, sensors and actuators distributed throughout a complex system involve different sampling rates, i.e., multirate sampling. The second reason rests on the fact that the use of multirate and periodically timevarying controllers can significantly improve the closed-loop performance of a sampled-data system in terms of model matching, sensitivity reduction, disturbance rejection, and pole and zero assignment with state feedback. This contribution is an in-depth treatment of these issues, and, as such, is also an essential element of this volume. The next contribution is "Maximizing the Fisher Information Matrix in Discrete-Time Systems," by Wendy L. Poston, Carey E. Priebe, and O. Thomas Holland. One of the most important aspects of the design and analysis problem for discrete-time systems is that of developing and verifying a material model of the system to which discrete-time control is being applied. One of the most important methods for model verification is the Fisher Information Matrix Technique. This contribution is an in-depth treatment of this technique, including illustrative examples for model development and verification. The next contribution is "Discrete-Time Constrained Linear Systems," by Jean-Claude Hennet. The existence of hard constraints on state and control variables often generate problems in the practical implementation of control laws. Methods for generating control techniques which avoid state or input (control) saturations and including these aspects in the system design are presented in this contribution. Numerous examples are presented throughout this contribution which illustrate the effectiveness of the techniques presented. The next contribution is "Digital Control with H~ Optimality Criteria," by Hannu T. Toivonen. The limitations of standard continuous and discrete design methods in the treatment of sampled-data control systems have recently led to the development of a robust control theory for sampled-data control systems. This contribution presents the various approaches to the development of robust control systems by means of solving the sampled-data H~ control problem. Several major new issues and techniques are also presented in this contribution. The next contribution is "Techniques in On-Line Performance Evaluation of Multiloop Digital Control Systems and Their Application," by Carol D. Wieseman, Vivek Mukhopadhyay, Sherwood Tiffany Hoadley, and Anthony S. Pototzky. This contribution develops a controller performance eval-
PREFACE
xi
uation (CPE) methodology to evaluate the performance of multivariable digital control systems. The power and utility of the method is exemplified in this contribution through its utilization and validation during the wind-tunnel testing of an aeroelastic model equipped with a digital flutter suppression controller. Through the CPE technique a wide range of sophisticated real-time analysis tools are available for rather complex discrete-time system problems. The final contribution to this volume is "Impulse Control of Piecewise Deterministic Systems," by Oswaldo L. V. Costa. There is a wide and diverse variety of discrete-time systems where control is taken by intervention; that is, the decision to act or apply control is taken at discrete times. In this contribution the impulse control problem of piecewise deterministic processes (PDPs) is addressed. Powerful computational techniques are presented and illustrated. The contributors to this volume are all to be highly commended for their contribution to this rather comprehensive treatment of discrete-time system analysis and design techniques. The contributors to this volume have produced a modern treatment of the subject which should provide a unique reference on the international scene for individuals working in many diverse areas for years to come.
This Page Intentionally Left Blank
7-/2-Optimal Control of Discrete-Time and Sampled-Data Systems Tongwen Chen Dept. of Electrical and C o m p u t e r Engineering University of Calgary Calgary, A l b e r t a C a n a d a T2N 1N4 Bruce A. Francis Dept. of Electrical Engineering University of Toronto Toronto, Ontario C a n a d a M5S 1A4
Abstract
This paper gives a complete state-space derivation of the discrete-time 7"/2-optimal controller. This derivation can be extended to treat a sampled-data 7-/2 control problem, resulting in a new direct solution to the sampled-data problem. A design example for a two-motor systems is included for illustration.
I.
Introduction
A recent trend in synthesizing sampled-data systems is to use the more natural continuous-time performance measures. This brought solutions to several new %/2-optimal sampled-data control problems [1, 2, 3], each reducing to an 7-/2 problem in discrete time. Discrete-time 7-/2 (LQG) theory was developed in the 1970's, see, e.g., [4, 5, 6, 7, 8, 9]. As in the continuous-time case, the discrete optimal controller is closely related to the solutions of two Riccati equations. In [10], the solution to a continuous-time 7-/2-optimal control problem was rederived using the state-space approach. This CONTROL AND DYNAMIC SYSTEMS, VOL. 71 Copyright 9 1995 by Academic Press, Inc. All rights of reproduction in any form reserved.
2
TONGWEN C H E N A N D B R L C E A. FRANCIS
gives a clean treatment of t h e problem and provides compact formulas for the optimal controller. Since complete, general formulas for t h e discrete optimal colltroller are not readily available in the literature, we ask t h e question here, can a state-space treatment be accomplished for discrete-time 'Ha problems? T h e goal in this paper is twofold: t o present a state-space solution t o the discrete 'Hz control problem and t o give direct formulas for an 'Ha-optimal sampled-data control problem with s t a t e feedback and disturl~ancefeedforward. Tllougli the results in the discrete-time case are known in various forms, we believe the derivation is new and quite self-contained, and therefore has some pedagogical value. Moreover, t h e formulas derived can be applied t o the sampled-data problem via tlle powerful lifting technique [11, 12, 13, 1.11. T h e organization of the paper is as follows. In the nest section we collect and prove some preliminary results on Riccati equations; t h e presentation follows closely t h a t in [lo] in conti~luoustime. Section 111 gives a complete state-space derivation of the discrete-time 'Hz-optimal control, first via s t a t e feedback and disturbance feedforward and then via dynamic output feedback. Section IV presents new direct formulas for a sampled-data If2 problem using s t a t e measurement. In Section V we apply t h e optimal sampled-data control in Section IV t o a two-motor system and compare with the optimal analog control. Finally, concluding remarks are contained in Section
VI. Tlie notation in this paper is quite standard: C is tlle complex plane, D C C is the open unit disk, and DD is tlle boundary of D, namely, the unit circle. Also, Z is the set of all integers and Z+ (Z-) is t h e nonnegative (negative) subset of Z. Tlle space C2(Z+), or simply C2, collsists of all square-sulnniable sequences. perhaps vector-valued, defined on Z+. Similarly for E2(Z) and E2(Z-). T h e discrete-time frequency-domain space 'H2(D), or simply 'H2, is the Hardy space defined on D. \.lie use R'H2 for the real-rational subspace of 'Hz. I11 discrete time. we use A-transforms instc)ad of ,--transforms, where A = z-I. If a linear discrete system G has a state-space realization ( A , B, C , D ) , then we denote the transfer rnatris D AC(I - AA)-'B
+
H,- OPTIMAL CONTROL
Finally, g"(X) stands for the transposed matrix g(l/X)'.
11. Riccat i Equation I t is well-known t h a t Riccati equations play an important role in t h e 'Hz optimization problem. T h e solution of a Riccati equation can be obtained via t h e stable eigenspace of tlie associated symplectic matrix if the s t a t e transition matrix of t h e plant is nonsingular. If this matrix is singular, as is the case when t h e plant lias a time delay, tlien the s.vmplectic matrix is 11ot defined; but we can use tlie stable generalized eigenspace of a certain matrix pair [9]. Let il, Q , R be real n x n matrices with Q and R symmetric. Define t h e ordered pair of 2n x 2n matrices
A pair of matrices of this form is called a synzplectic pair. ( T h i s definition is not t h e most general one.) Note t h a t if A is nonsingular, then H T ' H ~ is a symplectic matrix. Iritroduce the 2n x 2n matrix
I t is easily verified t h a t Ill J t I i = H 2 J l i a . Thus t h e generalized eigenvalues (including those a t infinity) for the matrix pair 11 (i.e., those numbers X satisfying I l l z = X H 2 x for some nonzero z ) are symmetric about the unit circle, i.e., X is a generalized eigenvalue iff 1 / X is [9]. Now we assun.~cN llas no ge~leralizedeigenvalues on 389. Then it must have it inside and 72 outside. Thus the two generalized eigenspaces & ( H ) and X , ( N ) , corresponding t o generalized eigenvalues inside and outside tlie unit circle respectively, both have dimension n. Let us focus on t h e stable subspace X , ( W ) . There exist
TONGWEN CHEN AND BRUCE A. FRANCIS
n x n matrices
X 1
and X2 such that A'i(H)-Im
[Xl] X2
"
Then for some stable n • n matrix Hi,
H~
[ ] [Xl] X1
- H2
X2
X2
(1)
Hi.
Some properties of the matrix X~X2 are useful. Lemma
1 Suppose H has no eigenvalues on OD. Then
(i) X{X2 is symmetric; (ii) X{X2 >_ O if R >_ O and Q >_ O. P r o o f R e w r i t e (1) as two equations"
(2)
AX1 - X I Hi + RX2Hi -- Q X 1 + X 2 -
(3)
A' X2 Hi.
Part (i) can be derived easily from these two equations (see, e.g., [15]). For part (ii), we define M "- XaX2 - X2X, and pre-multiply (2) by H[X; to get 9- /
H i'X 2"A X 1
"t
"
- H [ M H i + H ' i X 2"'R X 2 H i
9
(4)
Take transpose of ( 3 ) a n d then post-multiply by X1 to get
- X{QX1 + 3 1 - H[X(~AX1.
(5)
Thus equations (4) and (5) give I
~-I
-I
H [ M H i - M + HiX2RX2Hi + ~ I Q X 1 - O. This is a Lyapunov equation in M. solution is
Since Hi is stable, tile unique
(ND
AI - E tt~k(H[X~ RX2Hi + X{QXl)Hik, k=O
H~- OPTIMAL CONTROL
5
which is _> 0 since R and Q are _> 0. m
Now assume further that spaces
X 1
Xi(H), are complementary. Set X
"-
is nonsingular, i.e., the two sub-
Im
X2X11 9Then
Xi(H)-Im
[1] X
"
(6)
Note that the n x n matrix X is uniquely determined by the pair H (though X1 and X2 are not), that is, H ~ X is a function. We shall denote this function by Ric and write X = Ric(H). To recap, Ric is a function 7~2'~x2'~ ---, 7~'~xn that maps H to X, where X is defined by equation (6). The domain of Ric, denoted dora Ric, consists of all symplectic pairs H with two properties, namely, H has no generalized eigenvalues on 0D and the two subspaces
Im[~ are complementary. Some properties of X are given next. Lemma
2 Suppose H E dora Ric and X -
Ric(H). Then
(i) X is symmetric; (ii) X satisfies the algebraic Riccati equation A'X(I + RX)-~A-
X + Q - o;
(iii) (I + R X ) - A is stable. 1
P r o o f Setting X 1 - I and X2 - X gives (i) from Lemma 1; moreover, ( 2 ) a n d (3)simplify to the following two equations
A -(I + RX)Hi
(7)
6
TONGWEN CHEN AND BRUCE A. FRASCIS
- Q + X = A'XH,.
(8)
If A is nonsingular, so is Hi and then I + RX by (7); if A is singular, by [16] (Lemma 1.5) I RSy is still nonsingular. Hence
+
This proves (iii) since Hi is stable. Substitute ( 9 ) into (8) to get the Riccati equation.
Lemma 2 is quite standard, see, e.g.. [9. 1.51. Tlie following result gives verifiable conditiolls under wliich IZ belol~gsto do112 Ric.
Theorem 1 Suppose H has the fornz
with ( A ,B ) stabilizable and ( C . A) having n o unobservable nzodes o n dD. T h e n H E d o m Ric and Ric(H) > 0.
Proof We first show that H has no generalized eigenvalues on the unit circle. Suppose, on the contrary, that ejs is a generalized eigenvalue and
a corresponding eigenvector; that is.
Write as two equations and re-arrange:
, j e ( ~'
,-je)z = - c 1 c X .
Pre-multiply (10) and (11) by e-jez* and x' respectively to get
(11)
H,- OPTIMAL CONTROL
Take complex-conjugate of the latter equation to get _lrCxll
2
_ IIB',II 2.
=
Therefore B ' z - 0 and C x - O. So from (10) and (11) ( A - eJ~ ( A - eJ~
-
0
-
O.
We arrive at the equations z*[A - e j~
C
B]
-
0
x
-
O.
By controllability and observability of modes on OD it follows that x - z - O, a contradiction. Next, we will show that the two subspaces A'i(H),
[0]
Im
I
are complementary. As in the proof of Lemma 1 bring in matrices X1, X2, Hi to get equations (2) and (3), re-written as below (R = BB', Q = C'C): AX1
+ B B ' X 2 Hi
(12)
- C ' C X 1 -t- X 2 - A ' X 2 H i .
(13)
=
X 1 Hi
We want to show that X1 is nonsingular, i.e., Ker X1 - 0. First, it is claimed that Ker X1 is Hi-invariant. To prove this, let x E Ker X1. Pre-multiply (12) by x~H[X~ and post-multiply by x to get I Trl ~rl
x sli.~2X zHix +
X I
I
I
I
H i X 2 B B X 2 H i x - O.
Note that since X ~ X 1 >_ 0 (Lemma 1), both terms on the left are >_ 0. Thus B ' X 2 H i x - O. Now post-multiply (12) by x to get X I H i X - O, i.e., H i x E Ker X1. This proves the claim.
TONGWEN CHEN AND BRUCE A. FRANCIS
Now to prove that X 1 is nonsingular, suppose Oll the contrary that Ker X1 # 0. Then HilKer X1 has an eigenvalue, IL, and a corresponding eigenvector, x" Hix
I#1
-
#x,
O . II
This theorem has various forms ill the literature; for example, in [6] similar results were given when tile matrix A is nonsingular and in [9] an indirect proof was given that X1 is nonsingular. Our proof here is along the lines of a continuous-time proof in [17].
III.
Discrete-Time Case
This section rederives in a state-space approach the perhaps-known results for a discrete-time 7-/2-optimal control problem. We begin with the standard setup shown in Figure 1. We have used dotted lines for discrete signals and will reserve continuous lines for continuous signals. The input r is standard white noise zero mean, unit covariance matrix. The problem is to design a K that stabilizes G and minimizes the root-mean-square value of ~; it can be shown that this is equivalent to minimizing the norm on 7-12 of the transfer matrix from ~o to ~.
H~,-OPTIMALCONTROL
l
0a
G
13
V5 ............ ~
K
Figure 1" Tile s t a n d a r d discrete-time setup.
A.
State Feedback and D i s t u r b a n c e Feedforward
First we allow the controller to have full information. In this case, as we will see, the optimal controller is a constant state feedback with a disturbance feedforward. With the exogenous input being some pulse function, say, w = co0ae (r is a constant vector and ~Se the discrete unit pulse), we can even think of v as unconstrained. The precise problem is as follows: 9 Given the system equations
G"
~(k + 1) ~(1r
-
A~(k) + Bled(to ) + B2u(]r
--
C1~(1r
D11r
with tile assumptions
B2) is stabilizable; (ii) D~2D1:2-I and D12; (i) (A,
(iii) the m a t r i x A-A C1
D12
has full rank V,~ E 0D. 9 Solve the optimization problem
mi,~ I1~112 2"
vEg2e
+ D12v(k)
~
--
r
10
TONGWEN CHEN A N D BRUCE A. FRAUCIS
Note that for ease of presentation we initially allow v to be in 12,,the extended space for 12; however, the optimal v , to be seen later, will actually lie in e2. Assumptions (i) and (iii) are mild restrictions and (ii) basically means that the number of outputs to be controlled is no less than the number of control inputs and the control weighting is nonsingular. If D i 2 D I 2 is ion singular but not identity, we can normalize it by defining the new v to be ( D ~ ~ D ~ ~ ) ' / ~ V . The setup can be depicted as in Figure 2, where the transfer
Figure 2: The full-informatiol1 discrete-time setup.
We will first derive the solution for a special case and then come back t o the general one. 1. Orthogonal Case A11 additional assumption is now made: (iv) Di2C1 = 0. This assumption is an ortliogonality conditiol~: It amounts t o orthogonality of C1( and D12u i n tlie output 0.
Define the
H,- OPTIMAL CONTROL
and the transfer matrix
By Lemma 2, A
+ B 2 F is stable and so 9, E 723-12
Theorem 2 The unique optimal control is v,,t over, mill 11C112 = Iljcwol12.
= F(
+ Flw.
More-
In contrast with the full-information continuous-time ca.se where t h e optimal control is a consta.nt s t a t e feedba.ck, t h e discrete-time optimal control law involves a disturbance fecdforward term, and this is true even when D l l = 0. A useful trick is t o cha.~lgevaria.ble [lo]. Start with the system equa.tions
and define a new control va.riable
So in t h e frequency doma.in
where
icis as
above and
i; is seen
t o be
A t B2.F
gi(A) =
[ i : ; ]
T h e matrices j, and j, 1ia.ve tlie following two useful properties:
TONGWEN CHEN AND BRLCE A. FRANCIS
12
Lemma 3 The matrix grij, belongs to 'R'Hi and
Proof To simplify notation. def ne
Then we have the power series representations
Using these formulas, write j;ic as a series i n A, wit11 both positive and negative powers. It remains to check that tlle coeficie~ltsof XO,A, X2,. . . a r e all zero; this can be proved using the Riccati equation and the definitions of F and Fl. W The proof of the second statement is sin1il;ir.
Proof of Theorem 2 Since v is free in t2,, so is v. Thus we can formally write in the time domain
since 11 E t2, and by Lemma 3 G';G',wohdE L 2 ( 2 - ) . Then in the frequency domain we can write
Now note that irgi = I
+ BkXB2 by Lemma 3 to get
This equation gives the desired conclusion: The optimal fi is fi = 0 (i.e., v = F[ Flw) and the minimum norm of ( equals Ilbcwollz.
+
With vOptapplied, the resultallt system is stable since is stable; thus uOptindeed lies in C2, as commented before.
it
t B2F
H~,- OPTIMAL CONTROL
2.
13
General Case
Now we return to the situation at the start of Section III-A, without assumption (iv). Our approach is to reduce the problem via a change of variable to one where the orthogonality condition holds. Define a new control signal !
Ynew
-- Y +
(15)
D12C1~.
Note that Vnew is a free sequence in g2r if v is. The equivalent system, having { as its state vector too, is then shown in Figure 3, where
"..................
Gnew
Unew
Figure 3: The equivalent full-information setup.
A_B2D~2C 1
B1 Dla
B2 .] D12 "
The three assumptions made on G at the beginning of Section III-A are also satisfied by GnCw; for example, assumption (iii) is verified by the following matrix identity:
[A C1
i 0] [A D12
-D'12C1
I
(I - D12D]2)C1
D12
"
Moreover, G,~ew satisfies tile orthogonality condition D t l 2 [ ( / - D12D~12)C1] - O. Now invoke Theorem 2 to get the optimal Vnew, and then the optimal v via (15). Let us summarize. The given system is G:
~ ( k + 1)
=
A~(k)+ Bl~(k)+B2v(k),
~'(k)
=
C l ~ ( k ) + D11w(k)+ D12v(k)
~ =~O~d
14
TONGWEN CHEN A N D BRUCE A . FRASCIS
and t h e problem is min, IICI12. Under assumptions (i)-(iii), define
Theorem 3 The unique ol>timnl control is uOpt= F( over. mill I l i l l 2 = 114c~~112.
+ Flu.
Alore-
B. O u t p u t Feedback Now we study tlle 'Hz-optimal cont1.01 problem posed a t the s t a r t of Section 111, where the measured output $ does not have full information and therefore dynamic feedback is necessary. All discussion pertains t o t h e standard discrete-time setup. Let Tcwdenote t h e closed-loop system froin w t o i.i\Te say a causal, finite-dimensional, linear time-invariant controller h' is crdr-rzissible if it achieves internal stability. Our goal is t o find an admissible I< t o minimize Ilicwl12. Again, we will first d o the orthogonal case in detail and then present the solution for the general case. 1. Two Special Problems For later benefit, we begin with two special 'Hz-optimal control problems. T h e first speciul problem has a G of t h e form
with t h e assumptions
H~- OPTIMALCONTROL
(i) (A, B2) is stabilizable;
(ii) D ~ 2 [ C 1
D12
]- [0
(iii) the matrix
B2 ]
A-A C1
D12
has full rank VA r OD; (iv) A - B1C2 is stable. Since D21 - I, the disturbance, w, enters the measurement directly. Define A
H
X F
i~c( :~ )
0
I
- Ric(H) = - ( I + B~2XB2)-IB~2XA
-
-(I + B'2XB2)-~(B'2XB~ + DI2DII) [ A+B2F BI+B2F1 ] C1 + D12F
Dll + D12F1
"
Tile next result says that tile optimal controller achieves tile same performance as the optimal state feedback and disturbance feedforward were the state and the disturbance directly measured. T h e o r e m 4 The unique optimal controller is
~opt( A ) "-- [ A + B2 F - B2['1C2- B1C2 B1 + B2 F1 ] F~ ] F - F1C2 [ J~/~Oreo vc r~ m~n Ili_ 1,
was simulated and the responses are shown in Figure 10. Not sur.
250
.
.
.
.
200
.
9
.
"",,,,
.
.
, ,
150
lOO
50
0
0
0.05
o.1
0.~5
0.2
o.2_~
0:3
0;5
0:~
o.:~5
0.5
Figure 10: Optimal sampled-data controller without disturbance feedforward. prisingly, the response is very poor: The slave motor does not begin to move until the start of the second sampling period, by which time the tracking error is very large. VI.
Conclusion
Direct formulas for the sampled-data output-feedback case are not available because the lifted problem is inherently singular (/)21 - 0). This obstacle does not arise in the operator-theoretic approach of [1]. A c k n o w l e d g e m e n t The authors wish to thank P. P. Khargonekar and P. A. Iglesias for helpful discussions.
H 2- OPTIMAL CONTROL
33
References [1] T. Chen and B. A. Francis, "7-/2-optimal sampled-data control," IEEE Trans. Automat. Control, vol. 36, No. 4, pp. 387-397, 1991. [2] B. Bamieh and J. B. Pearson, "The 7-/2 problem for sampleddata systems," Systems and Control Letters, vol. 19, pp. 1-12, 1992. [3] P. P. Khargonekar and N. Sivashankar, "7-/2 optimal control for sampled-data systems," Systems and Control Letters, vol. 18, pp. 627-631, 1992. [4] M. Athans, "The role and use of the stochastic linear-quadraticGaussian problem in control system design," IEEE Trans. Automat. Control, vol. 16, pp. 529-552, 1971. [5] P. Dorato and A.H. Levis, "Optimal linear regulators: the discrete-time case," IEEE Trans. Automat. Control, vol. 16, pp. 613-620, 1971. [6] V. Ku~era, "The discrete Riccati equation of optimal control," Kybernetica, vol. 8, No. 5, pp. 430-447, 1972. [7] B. P. Molinari, "The stabilizing solution of the discrete algebraic Riccati equation," IEEE Trans. Automat. Control, vol. 20, No. 3, pp. 396-399, 1975. [8] B. D. O. Anderson and J. B. Moore, Optimal Filtering, PrenticeHall, Englewood Cliffs, N J, 1979. [9] T. Pappas, A. J. Laub, and N. R. Sandell, Jr., "On the numerical solution of the discrete-time algebraic Riccati equation," IEEE Trans. Automat. Control, vol. 25, No. 4, pp. 631-641, 1980. [10] J. C. Doyle, K. Glover, P. P. Khargonekar, and B. A. Francis, "State-space solutions to standard 7-12 and 7-/00 problems," IEEE Trans. Automat. Control, vol. 34, No. 8, pp. 831-847, 1989.
34
TONGWEN CHEN AND BRUCE A . FRANCIS
[ I l l 11. T. Toivonen, "Sampled-data control of continuous-time sys' , optima.lity criterion," A ~ t o n z ~ t i cvol. a , 28, No. tems with an H 1, pp. 45-54, 1992. 1121 Y. Yamamoto, "A new approacll t o sampled-data control systems - a function space approach method," Proc. CDC, 1990. [13] B. Bamieh and J. B. Pearson, "A general framework for linear periodic systems with application t o H ' , sampled-data control," IEEE Trans. Autonzat. Control, vol. 37, pp. 418-435, 1992. [I41 B. Bamieh, J. B. Pearson, B. A. Francis, a.nd A. Tannenbaum, "A Iifting technique for 1inea.r periodic systems with applicatiolls t o sampled-data control." Systems cind Control Letters, vol. 17, pp. 79-88, 1991. [15] P. A. Iglesias and I to,
the
and y(.) to interval (t0,t]
the for
functions
can
be
TECHNIQUES FOR REACHABILITY
represented by block vectors in the following manner:
Iu(to)
d(t o) 1 d(to,t ) = Ld(t_l )
U(to,t ) = Lu(t_l)l
..Y(to+ 1)
ILx(t)
x.(to + 1)1
X(to,t )
-
9
Y(to,t ) - Ly(t)
Using these notations we can write, output of the system"
for the state and the
x(t) = L(to,t ) X(to) + C(to,t ) U(to,t ) + G(to,t)d(to,t) X(to,t ) - L ( t o , t ) x(t 0) + M(to,t)U(to,t) + N ( t o , t ) d ( t o , t ) C L(to,t) X(to)C + C(to,t) U(to,t) + C G(to,t) d(to,t) Y(to,t ) = CL(to,t)X(to) + CM(to,t)U(to,t) + CN(to,t)d(to,t)
y(t)
-
where: C(to,t) - (At-to-lB ... AB B) G(to,t) - (At-to-ID ... A D D) B
AB
0... 0
B... 0
0
0
M(t0,t ) = A t -t 0-1 B ...
AB B
43
44
PAOLOD'ALESSANDROAND ELENADE SANTIS D AD
0... D...
0 0
0 0
N(to,t ) =
A t -t o- 1 D ... AD D and finally
L(to,t ) -
A(t-to )
.A L(to,t ) =
A(t_to )
It will be convenient
in the
sequel
to
denote
by U(t0,t)
the
set of all functions U(to,t ).
3.2 -
Constrained
systems
We now turn to the definition of constrained system. At a superficial level such definition is simply obtained associating an unconstrained system with a set of constraints for the system variablesl We should consider, however, the ensuing constrained system as a whole and completely distinct from the corresponding unconstrained system. And it turns out that, as we shall illustrate later on, the properties of this new system may be profoundly different from those of the original one. To begin with, the constrained system cannot be considered, in general, a linear one. A rather general form for the constraints is the following" f(to,tt, U(to,tf),d(to,tf),X(to,tf),Y(to,tf)) E Q(to,t f) V t 0, t f > t O
(3.2.1)
TECHNIQUES FOR REACHABILITY
45
where f is a function assuming values in some finite dimensional linear space and Q(t0,t f) is a given set in such a space. T i m e tf may well be 4-oo.
all,
We may consider the same constraints for some, instead of for pairs t 0, tf. However this is easy obtained letting Q(t0,t f)
be the whole space for those pairs t 0, t f,
for
which there are no
constraints. Thus a distinction between the two cases is not required. The set of these relations is called constraint system. Of course it is assumed that the constraint system defines non void sets of time functions for all variables and time intervals. Any function satisfying the constraint system is called admissible. In this chapter we are interested (with a few exceptions) to constraints bounding the only input, and, possibly, the disturbance of the system. Thus the above relation takes on the form: f(to,tf, U(to,tf),d(to,tf)) E Q(to,t f)
V t 0, tf~et o
(3.2.2)
We assume that such constraint system defines non void subsets of U t and D t V t o, which are respectively the set of admissible 0
0
input and noise functions. For simplicity and space reasons, we shall mostly consider the case in which no noise is present and give our general definitions accordingly. Problems involving disturbances are however of paramount importance. Later on, we shall devote Section 9 to one such problem. To give an example we introduce right away an interesting special case. Assume that the function g takes values in [Rm and that v(t0,tf) is a vector in such space. Then consider the following relation:
g(to,tf, U(to,tf)) ~ V(to,t f)
V to, t f > t 0
(3.2.3)
46
PAOLO D ' A L E S S A N D R O AND ELENA DE SANTIS
It is trivial to verify that this latter form can be reduced to the previous one. In fact it suffices to to define the set Q(t0,t f) as the cartesian product set X{(-oo,vi(t0,tf))" i = l , . . m } . Notice that this linear case becomes"
latter
form
of the
constraint
system
V t 0, t f ~ t ~
W(t0,t f) u(t0,tf)) ~; v(t0,t f)
in
the
(3.2.4)
where W(t0,t f) is a matrix with dimensions matching that of U and v. Despite the name we are far away from linear theory. The real nature of the problem is instead polyhedral. Anyway, it is still terminologically usual to call such constraint system a linear constraint system. We shall soon go back to this case in our first formal definition.
in
With reference to this form of constraints, there is no harm selecting a certain finite interval [t0,t f] and confining the
study to the response of a system in this single interval. This is precisely what many papers and books dealing with constrained systems do. In this case there would not be much to add at this point. However, the unconstrained system has a number of interesting properties, that depend substantially on the variability of parameters t o and tf. This suggest that our investigation be carried on in more depth. Some system properties are necessarily lost when we add the constraints, others may or may not be lost according to the features of the constraints, others may be altered and, finally, new properties may arise. In the basic properties of dynamicity, causality and stationarity [24] as well as in reachability itself the two time parameters t o and tf play a fundamental role. At a very basic level this role depends on the fact that for unconstrained case the function spaces relative to intervals of form [t0,t f) are strictly connected, since restricting e.g. input function obtained.
to
a
subinterval,
a
legitimate
input
function
the the an is
For these reasons we are lead to pose some conditions to be satisfied by the constraint system. The role of such conditions
TECHNIQUES FOR REACHABILITY
47
will become more and more apparent as our analysis develops. More precisely we consider the following conditions: (i) - Assume an admissible input u is defined on the interval [t0,tf), assume tf > t 0 + l and consider t i such that tf > t > t i
Then
both
the
restrictions
of
u
on
[t0,ti)
and
on
0"
It i,tf)
are
admissible inputs. (ii) defined, some
-
Assume
that
two
admissible
inputs
uI
on
the
intervals
[to,t i)
and
Then
the
function
t E [t0,ti)
and
u(t)=
respectively
tf > ti > t.0
u(t) = Ul(t )
if
u u2(t)
on if
and
u2
are
It i,tf)
for
f,fto,t~ defined ~
by
t E [ti,t ?
an
is
admissible input. (iii)
For
any
interval
[t0,t f)
the
identically
zero
function
is admissible. The first remark in order at this point is that assumption (iii) will be at time (and for a special case) weakened in such a way that, for the purposes of the question under study, the effect of the milder assumption is the same as that of the original one. We do not introduce any terminological distinction and stipulate that (iii) is in force whenever a substitute assumption is not explicitly stated. A constraints system satisfying the above assumption will be called a dynamical constraint system. We shall come back to this property in the definition below. Next we need to introduce the concept of stationary constraint system. For this purpose, for any given integer T, consider the shift operator S(t0,T ) defined on U(t0,tf) by:
(S(to,T)u)(t)= u(t-T)
u u e U(t0,t f)
(3.2.5)
t = t0+T,...,tf+T
The
operator
S(t o,T)
is
linear
and
invertible
and
maps
48
PAOLO D'ALESSANDRO AND ELENA DE SANTIS
U(t0,t f) onto U(t0+ T,tf+T). At this point we collect in the following definition a number of important concepts relating constraint systems and constrained systems.
DEFINITION
invariant) only
when
if
A constraint input function
1:
an the
input
function
system u(t0,t f)
is stationary is admissible
S(t0,T)u(t0,t f)
is
(or time when and
admissible
for
all
integer T. A constrained dynamical system is called stationary if both the corresponding dynamical system and the constraint system is stationary. A constrained system is called linear if both the dynamical system and the constraint system are linear. A dynamical system associated with a dynamical system of constraints is called a dynamical constrained system. In this definition we refer to properties of both the components of a constrained system for the sake of generality, but of course, we have already assumed, for the sake of simplicity, that the dynamical system is both linear and stationary. Notice, however, that we may still consider non stationary and/or non linear and /or non dynamical constrained systems. This is obtained by associating to our linear dynamical system a constraint system which is not stationary and/or linear and/or dynamical. Another interesting remark is that, to define a stationary constraint system it is not required that the constraint relation should be independent of t o and tf. For example consider the constraint system: t
X f u(t) t
< (tf-t0) v
(3.2.6)
0
As the reader will immediately of stationary linear constraint system.
verify this is a special case
At this point, before entering more in depth in the theory of constrained systems it is convenient to briefly recall some basic ideas underlying the classical unconstrained reachability theory.
TECHNIQUES
49
FOR REACHABILITY
3.3 - Review of unconstrained reachability theory
We initiate recalling the classical definition of reachability. Even though our dynamic system is stationary we refer reaehability to a given instant of time in preparation of the constrained case where, as explained above, time variant systems may well occur. DEFINITION 2 - With reference to a linear dynamic system, we call a point (state) z of Nn reachable at time t if there exists a 8
instant
of
[tr,t s)
such
time
tr,
that
t
the
r
< t
solution
initial condition x(tr) - - 0 at
time
is
called
t $,
i.e. the
s
and
x ( t$) -
reachable
an
x(.)
input
of
function
the
system
u
defined
in
corresponding
to
and to the input u assumes the value z z.
The
space
set of
at
t.
all
If
$
states
all
reachable
states
are
at
t$
reachable
at t the system is called reachable at t too. 8
$
Notice that the definition of reachability is characterized by two main facts. The first is that the initial condition is fixed to be the origin of the vector state space. The second is that reachability is a feature of possible evolutions of the system, occurring in the past, relative to the instant of reachability. An important observation, which provides a key argument in the analysis of reachability, is that, if a state z can be reached at t starting from zero at time t , then it can also be reached s
r
starting from zero at any instant of time, fact
it
suffices
the
system
x(t)--O, that
with and
r
to
steers the
apply initial
a
zero
input
condition
then to concatenate state from
zero
in
set
of
states
reachable
at
at t t
interval
to z at t .
[tp,tr)
to
In
other
words
s
starting s
the
In
x(t ) - O, thereby obtaining p this zero input with the input r
the
say tp, prior to t.r
(from
zero
state)
at
50
t
r
PAOLO D'ALESSANDRO
< t
$
is
contained
in
the
A N D E L E N A DE S A N T I S
set
of
states
reachable
at
the
same
time starting (from zero state) at earlier times. The set of states reachable from zero state starting at step of time ahead of t , starting at two steps of time ahead $
one etc.
form an increasing family of linear subspaces. Their union is the reachable space at time t . Because we obtain linear subspaces of $
a finite dimensional linear space, such sequence of linear subspaces can only increase up to a certain point and then it will become constant. In view of stationarity the set of reachable states does not depend on time and we can equate the first subspace to the space of states reachable at time one s t a r i n g from state zero at time zero, the second to the space of states reachable at time two starting from the zero state at time zero and so on. Clearly the first linear subspace is the range of the matrix B, the second is the range of the block matrix (B AB) and SO O n .
Finally combining the above observations with the Cayley Hamilton theorem we arrive to the conclusion that the set of reachable states is the linear subspace of the state space given by the range of the matrix C(O,n), which can be more simply denoted by
C(n). At times we state starting from is a simple link reachability from a at t from state x
are interested to the possibility of reaching a another state, different from the origin. There between reachability from the origin and state, say x. In fact we can reach the state z at t , if and only if we can reach from the
s
r
origin that state z-A(ts-tr ).
3.4 - Reachability concepts for constrained systems
Of course we might think to adopt the same definition of reachable state and reachable set at a certain time as in the unconstrained case. However, this would not be enough for constrained systems for a number of reasons.
TECHNIQUES
51
FOR REACHABILITY
We face now a radically different state of affairs. To mention a few novelties, the finite time reachability property does not hold anymore. What is reachable in f'mite time may be quite different from what is reachable in infinite time. Moreover the constrained system may not be stationary even though the unconstrained system is. Thus, for example, reachability ahead in time may be different from teachability from beforehand. Even the property that if we decrease the first extreme of the time interval the reachable set grows is missing if we allow for non dynamical constraint systems. Consequently a more refined definition is advisable in order to capture that greater complexity. The following definition formalizes a concept that already came to the fore in the arguments underlying unconstrained reachability theory.
DEFINITION 3 - A state z of a system (either unconstrained) is reachable at time t from time t g
t
r
in
t-
t
8
r
steps)
if
there
exists
an
r
input
(an
constrained or (or from time admissible
input
in the constrained case) such that the solution corresponding to initial condition x ( t ) = 0 and to such an input assumes the value r
x(t $)
=
z.
The
set of reachable
of reachable states from t $
states at time t 1, t $
g
is the union
of the
set
2 etc.
Note that such family of state space subsets is not in general an increasing one. Moreover if the state trajectory starting from zero state zero at t assumes a certain value w at an intermediate r
time
t
(t
< t. < t ) it is not necessarily true that the state w i r x s is reachable at time t from time t . Both these properties do i r instead hold in the case of dynamical constrained systems. The proof of this facts becomes trivial if one bears in mind the arguments on which were based our analysis of the unconstrained case. These observations enlighten the role of the dynamicity assumptions.
52
PAOLO D'ALESSANDRO AND ELENA DE SANTIS
In presence of constraints for the only input (and, possibly, disturbance), the previously illustrated link between reachability and reachability from a state holds good. This would not be true in general if the state were involved in the constraint system tOO.
This review of basic reachability properties does not exhaust all the interesting fields of investigation. Many other properties could be considered than there is space to cover here. However, we do treat the ease of approximate robust teachability in presence of noise. Appropriate definition and fundamental results will be given in an apposite section below.
3 . 5 - P o i n t w i s e in t i m e c o n s t r a i n t s
A simple but practically interesting particular form of the constraints is that of pointwise (in time) constraints for system's variables, where, at each instant of time, the value of the variable is forced to belong to a (nonvoid) set, that may be fixed or vary in time. Such constraints, for the case of the only input, are expressed by: u(t) E W(t)
Yt
(3.5.1)
where W(t) is a nonvoid set (called the constraining set, while W(.) will be called the constraining function). The function W(.) may well be a constant function. To represent absence of constraints at a certain instant t, it suffices to put W(t) equal to the whole space of input values. Clearly, for the constraints system to be dynamical it suffices that the origin belong to W(t) for any t. We shall weaken at times this assumption, in such a way, though, to surrogate the effects of the dynamicity hypotheses. Note also that the constrained system will be stationary if and only if W(.) is a constant function (whose value will be denoted by W). In the stationary case the reachable set will be denoted by R w whereas in the time varying case it will be denoted by Rw(t). The
symbol
R
without
subscript
will
denote
the
unconstrained
TECHNIQUES FOR REACHABILITY
53
system's reachable set. Obviously, whatever is W(.) (or W in the stationary case), Rw(t ) (or Rw) is contained in R.
4 - REACHABILITY UNDER GENERAL TIME-POINTWISE CONSTRAINTS Most of t~e results of this and the next section were stated in [22] and [23]'. Our first concern is to study how set operations on the constraining sets reflect on the reachable set. We can define any operation on the functions of the form W(.) by the corresponding operations on values of the functions. That is for example: (Wl(.) CI W2(.))(t ) - Wl(t ) N W2(t )
for any t
With this premise we can state the following theorem: THEOREM
1.
( i ) - If BWl(t ) c BW2(t ) for any t then RW ( t ) c
RW (t) for
1
2
any t. (ii) - Let A be a nonvoid set and
{W 9 . o~
E A} (briefly {W }) ot
be an arbitrary family of constraining functions. Then for any t: RN{ W } ( t ) c N {RW (t)} o~
Rt3 {W }(t) ~ t3 {S W (t)}. o~
any
(iii) - For any constraining function W(.) and real a, for finite family {Wt(.),...,Wk(.)} of constraining functions and
{al,...,ak } of reals, and for any t:
IMore specifically, Theorems 1, 2, 3, 4, 5 and Lemma 1 are adapted from [22], with kind permission from Pergamon Press Ltd, Headington Hill Hall, Oxford O X 3 0 B W , UK.
54
PAOLO D ' A L E S S A N D R O AND ELENA DE SANTIS
Raw(t) =
aRw(t ).
R~a.W.(t ) c E Ra.w.(t) 1
1
1
1
In the latter relation equality prevails if 0 E ABW.(t) u t,i. 1
PROOF -The first two statements have a straightforward so does the first statement in (iii). Therefore, if we for any W l and W 2
proof, and prove that
RW + W (t) c R W (t) + R W (t) 1
2
1
with equality prevailing
2
if 0 belongs to both ABWI(t) and ABW2(t)
for any t, then the rest of the proof will follow rather directly. Actually
if
z E R w + W (t) 1
starting
at some time
with Ul(t ) E Wl(t)
then
there
exists
an
input
u
2
t < t having the form
and u2(t) E W2(t),
solution x(.), with x ( t ) = 0 , that z belongs to the r.h.s, fact that the forced solution function.
u(t) = ul(t) + u2(t)
such that the corresponding
satisfies x(t) = z. At this point, set immediately follows from the is linear with respect to the input
Conversely if z E R W (t) + R W (t), that is z = z I + z 2 with 1
z I E RW ( t ) 1
2
with the constraint z1
starting
2
and z2 E R W (t), then let u I be the input (compatible
at
corresponding
defined
time
by W l) that
t I < t,
and
steers the
similarly
let
state from u2
be
the
0 to input
to z2, which will start at time t 2 < t. If t l = t 2 then
the control u l + u 2, compatible with the constraint defined by W I + W 2, will steer the system from the zero state at time t I to the state z at
time
t
in
view
of
the
linearity
of
the
solution.
Otherwise,
55
TECHNIQUES FOR REACHABILITY assume, an
without
input
such
u'
that
[tl,t2) ,
that
u'(t)E
which
is
restriction
of
coincides
with
Wl(t)
and
possible
in
will s t e e r the system from time t, and the proof is
generality,
that
u I in the
ABu'(t)=0 view
of
interval
for the
t I < t 2.
all
t
Consider
[t 2, t), in
hypothesis.
and
the
is
interval
Then
u'+u 2
state zero at time t I to the state z at
therefore
concluded.
Notice that W t c W 2 implies BW 1 c BW 2" We also remark that the inclusion relations in the statement ii of Theorem 1 may occur in proper sense, as will be shown by the following example"
EXAMPLE 1 Consider equations:
a
x l(t+l)
discrete
time
linear
system
described
by
the
= -mx2(t ) + u l(t) 0<mn- I
{~ AiB U}
(4.2)
i •O
This result has been generalized for the case of cone constrained inputs in [22]. For general time-pointwise constraints we can state the following theorem, in which we deal also with finite time teachability. To this effect we denote by R wk the set of states reachable from the origin in at most k steps of time.
T H E O R E M 2.
The set Rwk can be expressed as:
57
TECHNIQUES FOR REACHABILITY k-1
Rwk= Z A~B w
(4.3)
iffiO
Moreover, if 0 E A B W then k-I
Rwflim{ k->oo
Z AiB W i•O
}
(4.4)
The sequence of sets being actually an increasing sequence. Finally, if 0 E B W, then 0 G R w, A i B W c R w Vi, R w is invariant with respect to A and the set of all states from states belonging to R w is contained in R w + R w.
reachable
In the statement of this theorem the usual mathematical definition of the limit of an increasing sequence as union of the sequence itself is adopted. PROOF The first statement that requires a non trivial that of invariance. If x E R w, then for some k -
k X E ~ AiBW iffi0
Ax E
proof is
and hence
k A~, AiBW = i-0
k+1
~ AiBW i -1
but since 0 E BW
k+l
AiBW c
i-I of
k+l
~ AiBW c R
i-o
w
To prove the last statement decompose the response in the sum the free and forced response. The first, in view of the just
PAOLO D ' A L E S S A N D R O AND ELENA DE SANTIS
58
proved invariance, remains in R w. The same is true for the second by
definition
of
Rw
and
the
assumption
of
stationarity.
Thus
the
desired conclusion follows.
The statement of this theorem highlights some differences with the unconstrained case. The most noteworthy of these is that an unconstrained system has a finite time teachability property (that is, ff W = ~ n then the sequence increases at most up to the nth term), which does not hold in general. The properties
next natural question to ask regards of R w depend on the properties of W.
how do the In this respect
recall that an operator A is a contraction if IIAII s stipulate that A is a proper contraction if IIAII < 1 . premises we can state the following:
1. Let us With these
THEOREM 3. Assume still that 0 E A B W and that k is any positive integer then (i) - If BW is convex then both Rwk and R w are convex.
R
W
(ii) - If B W is bounded and A is a proper contraction then is bounded. If BW is unbounded then R is unbounded and hence Wk
such is R (iii)
W
-
If W has interior then both Rwn and R w have interior
relative to the subspace R. (iv) - If W is open then both Rwn and R w are open relative to So (v) - If B W is a subspace then both Rwk and R w are subspaces and Rwk = R w for any k ~ n .
Moreover R w is the minimal subspace,
that is invariant under A and contains the subspace BW.
TECHNIQUES FOR REACHABILITY
59
(vi) - If B W is a (convex) cone then both Rwk and R w are cones. Moreover R
is the minimal cone, that is invariant under A w and contains the cone BW. (vii) - If B W is a group under addition then both Rwk and R w are groups under addition. Moreover R
is the minimal subgroup of
w
~n that is invariant under A and contains the group BW
PROOF - For the sake of brevity we outline only a few crucial arguments for the proof. From these, from the previous results and from standard arguments used in linear reachability theory it is not difficult to build complete proofs. The proof of (i) follows immediately from the expression of the reachable sets given in Theorem 2 and elementary computation rules for convex sets (see e.g. [25]). As to (ii) note that, because BW is bounded, for some positive real r, BW c S r, where S r de~otes the c l o s ~ sphere about -I r -I . . the origin with radius r. Thus W c B S (where B is the reverse image function), so that, by Theorem 1, R W c RB-1Sr. On the other hand"
sup {llxll" x
R(B-1sr)}
--
k-I
= sup
{[1
E AiBu(k-i-1)[[
"u(i)
E
B-Isr} =
i =0 k-I
= sup { II E Aiz( k-i-1)ll
" z(i)
e
s r}
0
belong
to
the
non
W = { (Ul,U2): U l = - I , O~;u2~;1 } LJ { (Ul,U2): - l ~ ; U l ~ ; 0 , u2ffil } It is easy to verify that both Rwk, for every k > 1, and R w have interior; moreover, if some 0.25
k,
while
R
0.25 < m < 0.5 , Rwk may be not convex for is
w
convex
and
bounded.
In
fact,
if
< m < 0.5" Rwf{(xl,x2)'ll
x 2 ( t + l ) = -m x2(t) + u(t) with the constraint O~u(t)
Rw=
~g 1
Vt
{ (xl,x2) " xl = x 2}
This example shows that the condition (vi) in Theorem 3 is not necessary (the set W is a polytopr while the set R w is a subspace). Moreover, because a subspace is a convex cone, it also follows that the condition (vi) in the same theorem is not necessary. 1
EXAMPLE 5 Consider the system described by the equation" x(t+l)
= x(t) + u(t)
with input constraining set W = { O, 1 } for every t.
TECHNIQUES FOR REACHABILITY
The
reachable
set R w is the
group under addition. This Theorem 3 is not necessary,l
fact
63
set of all integers, shows
that
the
and
hence
a
(vii)
of
condition
A few further important remarks on the theorem are in order. As a special case note that if the system is reachable and W has interior then R w has interior, whereas if the system is not reachable
R
W
cannot
have
interior
even
if W
does,
since
R
W
is
contained in R. Moreover if W is convex and 0 E B W and R w is unbounded then R w contains a convex cone. Actually in this case R w is an unbounded convex set containing the origin and hence the recession cone of R is a nontrivial cone and is also the maximal in R w (see [25]).
W
convex cone contained
Note also that if a power of A is zero (that is, A is nilpotent) then any power with a greater exponent will be zero (exploiting the Jordan form of a matrix, it is not difficult to prove that the minimum power of A which is zero is at most the n-lth). Hence if a property of W is not inherited by the finite time horizon reachable sets, this also excludes that in general it is inherited by R w. Conversely it may happen that a property is not
in
general
inherited
by
Rw,
but
it
is
inherited
by
the
finite
time reachable sets. It is convenient to mention some obvious negative cases for which a property of W is not inherited by R w. This is the case when W is a closed set or when W is a sphere of a given norm or when W is a nonlinear manifold (actually in general neither the image under a linear map of a manifold nor the sum of two manifolds is a manifold).
64
PAOLO D'ALESSANDRO AND ELENA DE SANTIS
$ - R E A C H A B I L I T Y UNDER P O L Y H E D R A L C O N S T R A I N T S In this section, for simplicity, we make reference to properties of W, but it is clear that, as in the previous section, some generalization can be achieved making instead reference to BW. On the base of the first statement of Theorem 2, and the fact that a sum of polyhedra is a polyhedron, it is clear that in the present case Rv~ is a polyhedron. As is well known [26], if W is a polyhedron then (5.1)
W = P + L + C
where P is a polytope, L is a linear subspace and C is a pointed polyhedral cone. Regarding this decomposition we can state the following
LEMMA 1. The polytope P can be chosen to contain the origin if and only if W contains the origin. PROOF - In fact assume that W does not contain the origin. Then if P contains the origin it would follow that so does P + L + C , which is a contradiction. Conversely suppose that W contains the origin but P does not. Then because both {0} and P are contained in W, which is convex, it is possible to consider the decomposition: W = C({O} U P} + L + C where the polytope C({O} U P) contains the origin.II At this point if 0 E W, we can choose P according to l, and write, in view of Theorem 1
LEMMA
Rwk = Rpk + RLk + Rck
(5.2)
R
(5.3)
W
-R
P
+ R
L
+ R
C
65
TECHNIQUES FOR REACHABILITY
Therefore we can look separately at the cases where W is a linear subspace or a polyhedral cone or a polytope. The significant cases are those of a polytope and of a pointed polyhedral cone. Because a polytope is a finitely generated structure, it is not preserved by system's dynamics. However ff we consider the special case of finite time horizon reachability, which has foremost practical importance (in particular in optimization problems), then the polytopic structure is preserved: THEOREM 4. If W is a polytope then Rwk is a polytope for any k. PROOF - We know from Theorem 2 that k-I
=
Rwk
~,
AiBW
i-0
Each set in the sum is the image under a linear map of a polytope and hence is a polytope [25]. Moreover a finite sum of polytopes is a polytope (see again [25]), and therefore the desired conclusion has been achieved. I An upshot of the theory of systems over polytopes is the result below, which can be considered as the generalized bang-bang principle for discrete time systems. For a continuous time version of the bang-bang principle see e.g. [2]. Let
us
denote
by
{ej" j - 1 , . . , m }
(briefly
{e i})
the
set
of
extreme points of W and by E the set of all functions on {0,1,...} to {ej} These functions play the role of controls with bang-bang values. Then we can state the following" THEOREM 5. For any k the
extreme points of Rwk have the form:
k-I
Ak-i-lBu(i) iffiO
with u E E
PAOLO D'ALESSANDRO AND ELENA DE SANTIS
66
or, m other words, the set of extreme points of Rw~ is contained in the reachable set R
PROOF
- Since
W = C({ei}),
{ej} is the set of extreme
points of W,
it follows that AiB W = C(AiB{ei}).
dP
so that
From this fact
,dw
and the expression of R
Wk
it follows:
k-I
k-I
k-I
iffiO
i-O
iffiO
Rwk = Z AiBW =
Z C(AiB{ej}) = C( Z (AiB{ej}))
where in the last passage we have exploited the elementary result which ensures that, if A and B are arbitrary sets, then C(A+B) = C(A)+C(B). The desired conclusion is now an immediate consequence of the very definition of sum of sets. II
5.1 -
An o p t i m i z a t i o n e x a m p l e with an illustration of the bang bang principle
Optimization is not part of our concern here, but in this case an illustration of how the bang bang principles applies to optimization problems is just a few lines away. We consider here a simple functional and constraints that lead to an immediate solution by inspection. The structure of the solution will demonstrate the bang bang principle. Suppose that we want to maximize the functional Substituting the solution of the dynamic system we obtain: (f, x(T)) -- (f,C(T)u(T)) + (f, L(T)x(0))
(f, x(T).
(5.1.1)
The second term in the r.h.s, is constant and thus does intervene in the optimization. The first term can be rewritten as"
not
67
TECHNIQUES FOR REACHABILITY
(f, C(T)u(T)) Thus
if
we
= (C (T)f, u(T))
partition
the
vector
(5.1.2)
C (T) f
in
T
blocks
gi
(i = 0,..T-l) corresponding to those of the vector u(T) and bear in mind that the blocks of u(T) are independently constrained by u(i) E W it is clear that the problem diagonalizes into the T optimization problems" max (gi' u(i)) i -
subject to u(i) E W Next suppose that form (box constraints)"
the
m. :g u(i) ~
constraints
polytopical,
(5.1.3)
e.g.
i = 0,.., T-I
M.
1
be
1,..,T-I
1
of the
(5.1.4)
Then the optimum solution is clearly given by: mij u(i). = l
if
gij < 0
any value if
gij = 0
M..
gij > 0
tl
if
(5.1.5)
Because the maximum of the functional is surely attained on an extreme point of the reachable set at T and because the solution has the form contemplated by Theorem 5, this example confirms the bang bang principle. Finally we notice that the same arguments apply to the computation of the solution in the more general case in which the functional has the form: T i=l
where, of course
(f., x(i) = (f, x(T)) !
(5.1.6)
68
PAOLO D'ALESSANDRO AND ELENA DE SANTIS
f f
A practically verbatim repetition of the above steps leads to the diagonalization of the problem and to the solution. We leave the details to the reader for the sake of brevity.
6 - SYSTEMS O V E R CONES
We have already touched upon the case in which the input values are constrained to belong to a cone in Theorem 3, where we have introduced the theory of minimal invariant cones. If we introduce conical constraints for the other system variables too, then the theory extends in various directions. In [22], besides the above basic reachability result, conditioned and controlled invariant cones are introduced and their application to state constrained reachability theory is illustrated. Another interesting direction of investigation is that of positive systems (see e.g. [27] for the continuous time case). Some recent developments for the same case are in [28]. A generalization of the concept of positive system for discrete time systems (but the same concepts - if not the results - apply immediately to the continuous time case) is in [5]. Here constraints constraints.
we wish to complete the case of along lines that parallel the case
input conical of polyhedral
First of all we observe that any convex cone C in ~n is the s ~ n of its lineality subspace L plus a pointed cone given by L - n C. Because both of these two sets contain the origin, if we constraint the input to belong to a fixed cone at any time, then both the finite time reachable set and the reachable set decompose in the sum of the reachable sets corresponding to the subspace (which is a subspace) and that corresponding to the cone (which is a cone), according to Theorems 1 and 3.
69
TECHNIQUES FOR REACHABILITY
Next suppose that the input constraining cone C is polyhedral (that is, both a cone and a polyhedron at the same time). Notice that a linear subspace is a polyhedral cone and the intersection of two polyhedral cones is a p o l y h ~ r a l cone. Thus in the above decomposition the pointed cone L - N C is polyhedral too. Hence it is natural to complete our treatment examining the case of pointed polyhedral conical constraints. Incidentally notice that the nonnegative orthant of the space (i.e., the set of all vectors with non negative components) is a pointed polyhedral cone. Thus any theory of positive systems is a special case of the theory of systems over pointed polyhedral cones. A major fact regarding pointed polyhedral cones is that they are in a way the unbounded counterpart of polytopes. In fact in the same way as we can say that a polytope is the convex extension of the set of its extreme points, we can affirm that a pointed polyhedral cone is the convex extension of the union of its extreme rays. An extreme ray of a cone is a ray which is also a face of the cone. This result is, e.g., in [25]. For example the non-negative orthant is the convex extension of the coordinate axes. It may be more natural to use conical instead of convex extension. Thus let a minimal generating set be a set obtained taking a non-zero vector from each extreme ray of the cone. Then the cone is the conical extension of any minimal generating set. This similarity carries on, to reachable set. The following theorem 5.
same extent, to is the counterpart
finite time of Theorem
THEOREM 6. If
W
is
a
pointed
polyhedral
cone
then,
polyhedral (not necessarily pointed) cone and, generating set of W the cone Rwk has the form:
k-I
Co ( Z Ak'i-lBu(i)" u(i) e {ej}) iffiO
for
any
k,
Rwk
is
a
if {ej} is a minimum
PAOLO D'ALESSANDRO AND ELENA DE SANTIS
70
PROOF - Since W = Co({ej}), it follows that AiB W -
Co(AiB{ej}).
From this fact and the expression of Rwk it follows:
k-I
k-I
k-I
Rwk = ~ AiBW = ~ Co(AiB{ej})= Co(~ (AiB{cj})) i •0
i •0
i •0
where in the last passage we have exploited the elementary result which ensures that, if A and B are arbitrary sets, then Co(A +B) - Co(A)+Co(B). The desired conclusion is now an immediate consequence of the very definition of sum of sets. II We do not get involved here in optimization concepts though, because this would take us too far away.
7 -
CONSTRAINED STATE APPROACH TO THE CONSTRAINED INPUT REACHABILITY THEORY
In this section we develop a technique to compute the finite time reachable set of a given system, when the input constraining set is a polyhedron, in general time-varying. Here, to fix the ideas, we consider the set of states reachable at time T > 0 , starting from time 0. This set will be denoted, for simplicity, by ~r An argument similar to that at the beginning of Sec. 5 will immediately show that ~T is a polyhedron.
The theory is based on nonvoidness of a polyhedron. This respect to the bound vector of the polyhedron in question. (see [29]
a dual conical condition of condition is parameterized with the inequalities, which describe and [30]).
The idea is that of considering the unknown reachable set ~ T as a constraining set for the state at the same time T. By means of a backward recursion ([4],[5]), we find the description, at each step, of the set of the states admissible (that is, for which a
71
TECHNIQUES FOR REACHABILITY
solution exists) and with respect the z e r o vector polyhedron at time
with respect to this fictitious state constraint to the constraints on the input. By imposing that in the state space belongs to the admissible t - 0 , we arrive at giving the expression of ~r
It is important to stress that in our approach no assumption is required" neither on properties of the matrices of the linear system, nor on particular structures of the constraining sets.
Before needed.
illustrating
We shortly (i.e. the problem and the state are particularize the reachable set from
be t.,
this
some
preliminaries
are
describe the solution of the feasibility problem of existence of solutions), where both the input constrained to belong to given sets, and then we results, to solve the problem of finding the the origin, when the only input is constrained.
Consider the system (3.1.1) with D - 0 . For reasons that will soon apparent, it is also convenient to consider an initial time an initial state x(t.) = x, and a final time tf, with
1
0
method,
1
~g t i
~g t f
constrained
~g T .
in
a
We
assume
polyhedral
set
that
C
the t
for
state
all
of
t
the
in
system
the
is
interval
[0, T]: x(t) E C
t
V t E [0, T]
(7.1)
0 ~; t ~; T
(7.2)
or also, equivalently G(t)x(t) ~ M(t)
On the other hand the input of the system is constrained in a polyhedron W for all t in the interval [0, T-l]" t
u(t) E W or
t
V t E [0, T-l]
(7.3)
72
PAOLO D'ALESSANDRO AND ELENA DE SANTIS
F(t)u(t) < V(t)
DEFINITION 4: The problem state x at t i and to tf, if
on
defined
[ti, tf-1],
0 < t ~; T- 1
(7.4)
is feasible, there exists
U(t) E W t,
such
relative to an initial an input sequence u that
the
above
state
constraints arc satisfied by the solution of the system. We shall call any state, with respect to which the system is feasible, an admissible state, relative to the pair of times (ti, tf), and the given constraints. The computation of the set of admissible states is based on the following general backward recursion for the set of admissible states relative to initial times T - l , . . , 0 and to the final time T. Let D be the set of admissible states relative to (T-1,T) with T-1
respect to the constraints x(T) E C T , x(T-1) E ~n (that is, x(T-l) unconstrained) and u(T-1) E WT. I. Then it is clear that the set of states admissible relative to (T-2,T) constraints x(T) E Ca,, x(T-1) E CT_l, u(T-1) E W
T-1
and u(T-2) E W
T-2
with respect to the x(T-2) unconstrained,
is nothing but the set of states
admissible relative to (T-2,T-1) with respect to the constraints x(T-1) E ET_{ffiDT_{N CT_1, x(T-2) unconstrained and u(T-2) E WT_2. Generalizing
to
admissible states constraints x ( r ) E C r
arbitrary
t T 2, then C(TI)C (T2) > C(T2)C (T2). The result above readily implies that the system is reachable if and only if there exists a T such that the matrix C O ' ) C (1") is positive definite. The key of the argument lies, for necessity, in the fact that, because z is arbitrary, taking z = y, ~e a~rive to the conclusion that it must be (y, C(T)C ( T ) y ) > 1/p" Ilyll" p' ~so that there exists cr such that p > c7 > p'. Because G < p, z is no more reachable relative to the input bound norm G. Therefore, in view of Theorem 9, there exists a vector v such that:
(v, z) > a II C (T) vii